Learning from Expert-Corrected Misdiagnoses: A Self-Evolving Consultation Agent for Traditional Chinese Medicine after Failed Treatment

Abstract

In traditional Chinese medicine (TCM), treatment follows syndrome differentiation (bianzheng), the step that maps a patient's findings to a TCM syndrome. A misdiagnosed syndrome therefore leads directly to inappropriate treatment, and the error usually surfaces only after that treatment fails. The physician who sees the patient next must diagnose a condition already altered by the wrong therapy, identify the initial physician's error, and avoid continuing the failed therapeutic direction. We introduce TCM-MisDx, an interactive benchmark of 820 consultations after treatment failure built from three published monographs of TCM misdiagnosis cases. An agent gathers symptoms, tongue and pulse findings through the four diagnostic methods before reporting the current syndrome, the initial error, and a corrective treatment plan. We further present TCM-SEA, a self-evolving consultation agent that rewrites its own clinical skill document: failed consultations are compared with the expert physician's documented analysis, the agent proposes a minimal edit, and the edit is kept only if it passes an acceptance gate on held-in and held-out cases. The underlying language model is never updated. On the sealed test set, TCM-SEA raises the composite score from 61.8 to 65.5, syndrome accuracy from 53.9 to 57.2 and treatment-principle accuracy from 44.7 to 48.9, and lowers the therapeutic momentum rate, the fraction of corrective plans that keep the initial physician's failed direction, from 44.1 to 38.2. Learning from expert cases outperforms learning from outcome feedback alone, with the largest gains on the monograph of inappropriate-treatment cases. The two accepted edits are short, human-readable clinical checklists that physicians can audit directly.